Add quick performance benchmark tests

[DanRod1999]

Motivation

The existing performance regression tests require hub access, authentication, and downloading real iModels -- making them slow and hard to run locally for quick feedback. We need lightweight benchmarks that can run offline in seconds to catch performance regressions during development.

Approach

Adds a new pnpm test:quick-perf script in packages/performance-tests that generates synthetic iModels at runtime and benchmarks the transformer with minimal instrumentation overhead.

Key design decisions:

• Runtime-generated iModels -- no auth, no hub, no .env required. Creates a StandaloneDb with 15 custom ECSchemas and 10k PhysicalObject elements.
• BenchmarkTransformer subclass -- overrides processSchemas() and process() to capture phase-level wall-clock timing, plus counts elements/relationships/schemas via lightweight counter-only overrides (no per-call performance.now() to avoid measurement overhead).
• ECReferenceTypesCache benchmark -- standalone test that directly times initAllSchemasInIModel() since those metrics are useful but don't belong as logging in the library itself.
• Console output -- clean table with schemas exported, processSchemas time, elements exported, process time, and avg per element.

Sample output

  Quick Perf Benchmark Results
  ──────────────────────────────────────────────────
    Schemas exported:       15
    processSchemas():       792.18 ms
  ──────────────────────────────────────────────────
    Elements exported:      10005
    Relationships exported: 0
    process():              19004.90 ms
    Avg per element:        1.8995 ms
  ──────────────────────────────────────────────────

Notes

• The .cache directory generated by IModelHost.startup() is now gitignored.
• A VS Code launch config is included for debugging.
• Avg per element (~1.9ms) is within ~7% of the existing TransformerPerf.test.ts (1.78ms), with the gap explained by the 15 additional schemas adding metadata overhead.

Closes #328